Model Card for Gemma-3-270M-IT (Aria Quant Bundle, q4)

Model Details

Model Description

• Developed by: Google
• Quantized and distributed by: Aria Compute
• Model type: Dense Transformer decoder-only (language, instruction-tuned)
• Language(s): English (primary), Chinese, and 30+ additional languages
• License: Apache 2.0
• Finetuned from model: google/gemma-3-270m-it

Model Sources

• Original Repository: google/gemma-3
• Original Paper: Gemma 3 Technical Report
• Aria Compute Dashboard: ariacompute.com/dashboard/models
• Aria Engine: ariacompute.com

Uses

Direct Use

• On-device chat and conversational assistants
• Real-time text completion and basic code snippet generation
• Instruction-following tasks for mobile and IoT applications
• Lightweight text embeddings for on-device retrieval and classification
• Short-form summarization of notifications, messages, and local content
• Local document analysis up to 32K context (chunked)

Target Devices

Note: KV cache is extremely compact due to aggressive GQA (18 layers × 2 KV heads × head_dim=64), ~20% of an equivalent full-attention model's cache, enabling practical 32K context on sub-1 GB devices.

Out-of-Scope Use

• Long-form creative writing (>2K tokens per generation)
• Mathematical theorem proving or formal verification
• Full program/application synthesis
• Multimodal input (this model is text-only)
• Real-time audio/speech processing (use Aria speech models)
• Safety-critical decision systems without human oversight
• Tasks requiring factual precision beyond the model's ~270M parameter capacity

How to Get Started with the Model

Download from Aria Compute

Quantization Recipe

• Bundle size: ~0.13 GB (BF16 original: ~0.54 GB)
• Generation quality: Awaiting gen_quant_eval audit. Per-channel quantization preserves per-output-channel distribution characteristics. Formal quality benchmarks against FP16 and other Aria quant recipes are pending
• Calibration-free: Hadamard pre-processing + per-channel quantization, no task-specific calibration data required
• Other precision options: Also available for Gemma-3-270M-IT: gemma-3-270m-it_q8_channel (per-channel 8-bit) and gemma-3-270m-it_q326_channel (mixed precision, attn 4-bit + FFN ~3-bit)

Model Architecture

• GQA (Grouped Query Attention): 2 KV heads serving 10 query heads (group 5) — KV Cache memory is 5× smaller than full attention
• GeGLU activation: GELU with tanh approximation gating, efficient for on-device inference
• RoPE position encoding: Standard 10K base frequency supporting 32K context length
• RMSNorm pre-normalization: Lightweight normalization before each sub-layer

• ~75% smaller parameter count (270M vs 1.1B) — ~4× reduction, the smallest Gemma 3 variant
• Fewer layers (18 vs 26) with narrower hidden dimension (640 vs 1,152) and narrower FFN (2,560 vs 4,608)
• Halved head dimension (64 vs 128) — lighter per-head computation compensated with more query heads (10 vs 9)
• More aggressive GQA (group 5 vs group 3) — further KV cache compression, critical for fitting 32K context on memory-constrained devices
• Same 32K context length — full long-context support preserved despite the smaller parameter budget
• Instruction-tuned variant — this bundle is based on the IT (instruction-tuned) checkpoint, not the base pre-trained model, providing better instruction-following capability out of the box

Bias, Risks, and Limitations

Limitations

• Reasoning depth: Multi-step logical reasoning is highly limited for a 270M-class model. Verify outputs in high-stakes scenarios; consider larger Gemma 3 variants for reasoning tasks.
• Mathematics: Simple arithmetic may be attempted but is unreliable. Advanced quantitative reasoning is out of scope. Use larger models for mathematical tasks.
• Code generation: Capable of single-line completions and basic snippets; unreliable for multi-line code or structured programs.
• Factual knowledge: Very limited world knowledge due to ~270M parameter scale. Always verify factual claims against authoritative sources. This model is best suited for instruction-following and lightweight text processing rather than encyclopedic knowledge retrieval.
• Instruction following: Handles simple single-constraint instructions. Complex multi-constraint prompts may cause significant degradation, especially at longer contexts.
• Quantization drift: 4-bit per-channel quantization may exhibit generation drift versus FP16, especially on ambiguous or open-ended prompts. For higher fidelity, use gemma-3-270m-it_q326_channel (mixed precision, attn 4-bit + FFN ~3-bit) or gemma-3-270m-it_q8_channel (per-channel 8-bit).

Bias and Risks

• Bias: As with all large language models trained on web-scale data, Gemma 3 may reflect societal biases present in its training corpus. Evaluate outputs before deployment in sensitive domains (hiring, healthcare, law).
• Toxicity: The instruction-tuned model has been safety-aligned with RLHF. However, no safety filter is exhaustive. Consider an additional output classifier in high-risk environments.
• Hallucination: May generate plausible-sounding but factually incorrect information. Implement output verification for critical applications. Hallucination risk is elevated for smaller models due to limited memorization capacity.
• Dual-use risk: Text-generation capabilities could be misused for spam, disinformation, or impersonation. Deploy responsibly and in accordance with the Apache 2.0 license terms.

Recommendations

• Adding a lightweight output safety classifier for user-facing deployments
• Verifying factual claims with external knowledge bases
• Not using the model for high-stakes decisions without human review
• Considering larger Gemma 3 variants (1B or 4B) for tasks requiring stronger reasoning or factual recall